LVALH: An Image Hashing Method Based on Lightweight ViT and Asymmetric Learning

Most current deep supervised image hashing methods depend on Convolutional Neural Network (CNN) and symmetric learning strategy for feature extraction and model optimization. This dependency relationship often leads to insufficient global information in the derived image features and limited model generalization capabilities, subsequently impacting the quality of hash codes generation and the performance of hash retrieval. To address these issues, this paper proposes a novel supervised lightweight image hashing method that utilizes a lightweight Vision Transformer (ViT) model coupled with asymmetric learning strategy. This refined ViT model synergizes the strengths of CNN and Transformer structures, enabling it to effectively capture both local and global image features, and strike a balance between computational efficiency and enhanced feature extraction. Then, the asymmetric learning strategy leverages supervised information from both training set and database to formulate an asymmetric loss function, which enhances the model's convergence and optimization, significantly improving its generalization capability. Finally, extensive experiments are conducted on two public datasets, CIFAR-10 and NUS-WIDE. The results indicate that the proposed method outperforms other state-of-the-art methods.